Fred Weber, chief technology officer at Advanced Micro Devices, sat down with Electronic News to talk about the future of processing, AMD’s shifting business model and the need for multi-core chips. What follows are excerpts of that conversation.

Electronic News: Does the PC-centric approach still work for AMD, or does the company now have to change its approach to look at the consumer market?
Weber: We’ve come from the strength and capability of the PC. The PC moved to the server. Intel did that, and we did that in a very big way with the Opteron. Yes, there’s a PC starting point, but there’s also a server starting point. A lot of AMD’s success is that we brought both of those sensibilities. We have a long history with x86 PC processors. The people who led us into the server market, myself included, came from places like Digital Equipment, Encore Computer and Hewlett-Packard. This was a world of Risc, high-performance, server computers and system-level design. We grew up with [Digital Equipment Corp., VAX creator] Gordon Bell’s rule of thumb that the cost of the computer is proportional to the power supply. That’s not something that people from the PC world ever think about. Whenever we take the PC into a new area, you have to immerse yourself in what that new area needs and how the PC can serve it. The same is true in consumer electronics. It’s a fool’s errand to take the PC into consumer electronics. But to take many of the virtues of the x86 processor, the PC open architecture and the buses that have been developed there, and build a consumer electronics solution that uses those underpinnings can be very valuable because of the universality of software. In consumer electronics, cost flexibility is most important. In the PC space, performance was the most important.

Electronic News: How do you break into that space, though?
Weber: By leveraging partnerships with companies that have been in that space for a long time.

Electronic News: Are these US-based or global partnerships?
Weber: They’re all over the world. And these are partnerships in the broadest sense. They may be customers we work closely with, where they would specify what they want.

Electronic News: Looking at AMD five years out, who is your top competition going to be? Will it still be Intel?
Weber: There’s a set of people who are going after the semiconductor aspects of consumer electronics. Intel is one of them. Motorola, Samsung and TI are players in that space, too. Moving into consumer electronics means less Intel dominance, though.

Electronic News: What will most people be using five years from now? Will it still be a PC or will it be some sort of handheld device?
Weber: For me, the PC is too confining. I believe that peer-to-peer networking is a more natural model than a client/server model. The idea of one big device with lots of dumb devices around it is not a good model. That doesn’t mean there aren’t smarter and dumber devices, some more capable and some less capable. But you do better to think of them all as fairly capable agents in a world of interaction. If you look at the history of computing, we go back and forth between centralised and decentralised as the clients become more powerful or too complicated. If you make the devices peers, it’s amenable to a client/server model and complete independence. It’s great to be on the network, but sometimes you’re not on the network and need data stored locally. That’s why the idea of a thin client with no native capability is weak.

Electronic News: So where do you start?
Weber: I start with glass-LCD screens. There’s big glass and small glass, and it all deserves a fair amount of compute power, a software stack, connectivity and a certain amount of storage. Whether it’s a PC, which is a fairly big piece of glass connected to electronics, or a TV, which is a very big piece of glass, or it’s a cell phone, which is a small piece of glass, they all will have display capability, connectivity, storage and a deep software stack. They’re not all PCs, but they are all computers to organise and display the world.

Electronic News: So they all have a computer architecture?
Weber: It’s called computer architecture for a reason. The design of a building, which is classical architecture, is a merging of a deep knowledge of materials-what is possible, art, and what the customer needs. You meld the fine arts, the craft and the building capability into a solution. That is what we do. We merge what Moore’s Law and other laws of physics and manufacturing make possible, what looks good and what is needed. You put all this together in a solution. We think that in the next decade, one of the most valuable things a semiconductor company can do is work closely with its customers to make sure they’re getting what they need. What’s different is that this is architected, customer-centric innovation. You take the head of design to work with the customers to understand what they’re doing and essentially co-create something. Two years out we say we’re thinking about this and this and you find out what resonates with them. We also try to understand what problems they’re having.

Electronic News: What becomes the bottleneck of the future?
Weber: Software.

Electronic News: The operating system or the applications?
Weber: The stack is the problem. You can think of this like the population of the world growing exponentially. When that happens, you get more and more layers of government. You go from a city to a county to a state to a country to multinational organisations. The reason for all of this layering is that you need to limit the amount of interface. To serve an ever-bigger population, you always need another layer above it to tie it all together. Every time you add a layer, you add a layer of complexity about who does what and who’s responsible for what. Software layering does the same thing. You get another layer every time we get more performance because you need to do something with that performance. It adds an abstraction layer, so you go from DOS to fully functional OSes to Codec layers to middleware and presentation layers. As it gets taller and taller, there’s interaction between the layers and a systems integration of software becomes a harder and harder problem. The bottleneck of the future is developing the specific software you need and the systems integration.

Electronic News: How about the network?
Weber: It’s always a bottleneck. But there are three tricks in our bag. One is communication - moving data from one place to another. Another is storage, which is the ability to keep some data locally. And the third is computation. That’s used to transform or compress data. Whenever you find communication is too expensive or too difficult, you start to attack that problem. You may get more bandwidth, or you can put in a bigger disk and do more caching. There’s the ability to do video on demand, where you say you want to watch something at a certain time and it’s sent across the network to you. TiVo, meanwhile, is a caching approach. It says the bandwidth is limited so you locally grab it. The network may slow things down, but it’s actually one of the enablers rather than a real bottleneck.

Electronic News: Is AMD staying on the Moore’s Law road map?
Weber: If you look at the basis of Moore’s Law, which is the number of transistors, then for at least the next five years, and probably 10 or more, we do see a doubling of the density of transistors ever two years or so. That will happen for at least three more generations, and probably several more generations after that. People have been betting against that continuation for a long time and have always been wrong.

Electronic News: Is it really that they’re betting against it, or is it a matter of doing other things that aren’t on the road map?
Weber: First of all, Moore’s Law is going to be possible. There are two other things going on. For the past 20 years, that has always translated into higher single-processor performance. That is going to slow down dramatically. We’ve taken those transistors and done what we can do to make faster and faster processors. We will still make faster processors, but the rate of performance increase will slow down dramatically. Doubling the number of transistors now increases performance by about 20 per cent. If instead you go to two processors, it may go up 60 per cent or 70 per cent. It’s always better to have more single-threaded performance than two. If I could get 45 per cent single processing or 60 per cent dual processing, I’d take the 45 per cent.

Electronic News: Because it’s easier to write the software?
Weber: Yes. But if the discrepancy is 20 per cent versus 60 per cent, I’d probably take the 60 per cent. That’s enough to get you over the hurdle. By going to multi-core, we’re going to stay consistent with what we’re used to.

Electronic News: Do you get an advantage out of that versus two or three independent processors?
Weber: You do get some advantages. You can share the on-chip caching, the memory interfaces-and the pins are very expensive.

Electronic News: Is this like combining two companies and keeping one HR and one accounting department?
Weber: It’s a little like that, but it really does have long tentacles. There’s no room for another chip on many boards. It takes more space, additional heat sinks, and the economics are problematic. Why has a dual processor system never taken off on the desktop? It’s a cost-benefit issue. Single processors keep getting faster and faster. To burden every PC motherboard with a space to put a second processor would add $10 to $15 to the cost of every PC, and it wouldn’t be used very often. Then it’s a self-fulfilling prophecy, so if you don’t mass produce the boards then they cost $100 extra because they’re rare. When we put two cores on one chip we solved that problem. A single-core chip or a dual-core chip both plug into the same motherboard.

Electronic News: But isn’t a dual-core chip harder to manufacture?
Weber: From a yield point of view, there is some logic to going with two separate chips. But cost is still lower with smaller chips. Once I have four cores on a chip, it’s very likely that some chips will have one core broken, but very few will have more than one core broken. So we’ll have a huge number of chips with three cores working, which is perfectly fine. Redundancy has some nice properties.

Electronic News: What else affects performance?
Weber: One is multiple cores. The second is different needs. If you’re trying to build a phone, you care about how big the chip is and how much power it uses much more than you care about performance. We’re integrating other functions onto a chip.

Electronic News: Will there be one device that you take with you from the office to the road to your home that works as a phone and a PC?
Weber: I think there will be. It’s difficult to predict when, however. It’s a very compelling model in a lot of ways. There are a lot of reasons why it doesn’t quite work today - screen size, for example, and what do you do about a disk drive? But at the same time, carrying a lot of devices doesn’t make sense. Your personal data may be out on the network, but you’re not always connected to the network. It’s too compelling a model, and it’s clearly possible. It will happen, but when is a very hard thing to say.

www.amd.com